A Bayesian Graphical Model for ChIP-Seq Data on Histone Modifications Riten MITRA1, Peter MULLER¨ 2∗, Shoudan LIANG3, Lu YUE4,YuanJI5∗ 1 ICES, University of Texas at Austin 2 Department of Mathematics, University of Texas at Austin 3 Department of Bioinformatics and Computational Biology, The University of Texas, MD Anderson Cancer Center 4 Department of Leukemia, The University of Texas, MD Anderson Cancer Center 5 Center for Clinical and Research Informatics, NorthShore University HealthSystem August 10, 2012 ∗To whom correspondence should be addressed:[email protected], [email protected] Abstract Histone modifications (HMs) are an important post-translational feature. Different types of HMs are believed to co-exist and co-regulate biological processes such as gene expression, and therefore are intrinsically dependent on each other. We develop inference for this complex biological network of HMs based on a graphical model using ChIP-Seq data. A critical computational hurdle in the inference for the proposed graphical model is the evaluation of a normalization constant in an autologistic model that builds on the graphical model. We tackle the problem by Monte Carlo evaluation of ratios of normalization constants. We carry out a set of simulations to validate the proposed approach and to compare it with a standard approach using Bayesian networks. We report inference on HM dependence in a case study with ChIP-Seq data from a next-generation sequencing experiment. An important feature of our approach is that we can report coherent probabilities and estimates related to any event or parameter of interest, including honest uncertainties. Posterior inference is obtained from a joint probability model on latent indicators for the recorded HMs. We illustrate this in the motivating case study. An R package including an implementation of posterior simulation in C is available. KEYWORDS: Auto logistic regression; Epigenetics; Histone modifications; Markov chain Monte Carlo; Markov random fields; Network model; Pathway dependence. 1INTRODUCTION Genes are expressed in eukaryotes through complex processes that involve multiple layers of control. In addition to transcription factors that activate or repress their target genes by binding to DNA response elements in the promoters of the genes, the eukaryotic cells have additional layers of control of gene expression by chemical modifications to the his- tones. Histones are small proteins; they form a core around which DNA is wrapped, forming 1 nucleosomes. The combination of DNA and nucleosomes is called chromatin (Figure 1). Nucleosomes are the basic in-vivo structural unit of DNA, similar to thread wrapped around a stool. The thread is 147 base pairs of DNA and the stool consist of an octamer of the four core histones (two sets of H2A, H2B, H3, and H4). While the role of H2A and H2B are primarily structural, histones H3 and H4 play key roles in integrating a variety of signals that regulate gene transcription. In particular, these two histone proteins have “tails” — strings of amino acids that protrude outside of the basic nucleosomal structure and make contact with DNA. Specific post-translational modifications of the amino acids in these hi- stone tails (e.g. methylation, acetylation, phosphorylation, sumoylation, and ubiquination) interact with other proteins to either relax the chromatin and promote transcription, or to condense the chromatin into a closed form, which excludes transcription factors and result in gene silencing. Majority of these histone modifications (HMs) are epigenetic, i.e. that they are faithfully preserved in DNA replications. Thus daughter cells inherit the same patterns of HMs that occur in the parent cell. A comprehensive list of histone modifications appears at http://bioinfo.hrbmu.edu.cn/hhmd (Zhang et al., 2010). The correlation of HMs with translational activity and occurrence of promoters has been been well documented in Bernstein et al. (2002), Kim et al. (2005) and Roh et al. (2005). For example, histone methylations influence gene activation and repression, and histone acetylations are associated with a variety of functions such as gene activation, nucleosome assembly, higher-order chromatin packing and interactions with non-histone proteins (Grant and Berger, 1999). HMs also play an influential role in DNA damage repair and chromosomal segregation. See Bergink et al. (2006) and Andersson et al. (2009) for details. Biochemically, different enzymes can facilitate different modifications to histones in par- allel, thus making it possible for the co-existence of multiple HMs. It is hypothesized that the combination of different HMs dictates the status of a gene being permissible to tran- scription (active) or not (repressed) (Strahl and Allis, 2000). The hypothesis is known as 2 the histone code. Understanding the histone code is among the most exciting challenges facing scientists in genomics research today. Motivated by the histone code hypothesis, researchers have attempted to demarcate func- tional domains over the genome by signatures of histone patterns (Liu et al., 2005; Pokholok et al., 2005; Heintzman et al., 2009) and reported some interesting findings. However, a full description of the histone code remains elusive until now. Attempts to decode the combina- tion of HMs using quantitative methods have been primarily based on descriptive statistics, and independent and HM-specific hypothesis tests. For example, Wang et al. (2008) tested the enrichment of 39 individual HMs across the genome and identified a set of 17 HMs with largest p-values based on a Poisson assumption for the HM counts from a ChIP-Seq experi- ment. They used a Bonferroni correction to adjust for the multiple comparisons. They also computed pairwise correlations between any two HMs and conjectured that the set of these 17 HMs serves as a backbone of the dynamics of the chromatin structure genome wide and plays an important role in gene regulation. Building on these results, we propose to apply graphical models to investigate the com- plex dependence relationship of multiple HMs. To our knowledge, this is the first attempt to characterize the interaction of the HMs based on probabilistic graphical models. These models will allow us to identify the type and the strength of interaction of co-existing HMs. More importantly, the models provide a full probabilistic description of any subgraph in- volving subsets of HMs. For example, we present in Section 7.1 a highly connected graph involving the 17 backbone HMs. Our results confirm the findings in Wang et al. (2008) and elaborate their results with a full descriptions of all interactions. In addition, applying the same graphical model to the full list of 39 HMs in Wang et al. we reveal new insights in the entire HM-interaction map, thus making a step forward towards filling in part of the puzzle of the histone code. For the 17 backbone HMs, we shall focus only on dependence across HM’s, assuming independence across loci. In the full data analysis we will restrict inference 3 to coding regions only, thus reducing possible spatial correlations. In the next section we give a brief overview of a ChIP-Seq experiment that was carried out to record the HMs, and we discuss the statistical challenges posed by such data. In Section 3 we introduce the concept of Markov random fields (MRF) as a tool for describing dependence. Section 4 elaborates on the proposed probability model and Section 5 describes posterior inference. In Section 6 we present a simulation experiment to evaluate the performance of the proposed graphical model. In Section 7 we report results of the analysis of the ChIP-Seq data. Finally, we conclude with a discussion in Section 8. 2 ChIP-SEQ DATA We analyze data from a ChIP-seq experiment for CD4+ T lymphocytes (Barski et al., 2007; Wang et al., 2008). ChIP-Seq is a new technique that combines chromatin immunoprecipita- tion (ChIP) with massively parallel DNA sequencing (Seq) to identify genome-wide binding patterns of DNA-associated proteins, such as HMs. In such an experiment, HM-specific an- tibodies are used to preferentially extract the histones along with the DNA associated with it. DNA that cross-linked with the specific HM is then randomly broken up, by sonication, into pieces of several hundred nucleotides. The DNA pieces are used to generate short reads, which are sequenced and mapped to the genome. At the end of a ChIP-Seq experiment, the final measurement is the counts of DNA short reads that are mapped to specific loci on the genome. A loci here refers to a DNA segment of around 2,000 base pairs. The larger the DNA short reads j count, the higher the amount of the HM. By using 39 different antibod- ies and carrying out a ChIP-Seq experiment for each HM, Barski et al. (2007) and Wang et al. (2008) report HMs counts for m = 39 types of HMs including 18 acetylations, 20 methylations, and one special histone modification H2A.Z. We keep all genomic locations with at least one enriched HM and drop all the windows 4 where none of the 39 HMs are enriched. For the purpose of this screening, we use the peak-calling program SICER (Zang et al. 2009) to decide enrichment. SICER parameters were set to W SIZE=200, GAP SIZE=600, EVALUE=1000, FRAG SIZE=150. Also any adjacent windows with unchanged SICER calls for the 39 HM counts were merged to create larger regions. The data is arranged in an m × n matrix Y =[yit] of counts. Each row, i =1,...,m represents a type of HM, and each column, t =1,...,n, represents a genomic location. That is, the count yit reports the number of HMs of type i at location t. In Sections 7.1 and 7.2, we analyze two datasets with m = 17 or 39, and n =50, 000 or 33, 681, respectively.